Fare Shift: A Taxi Shift on Real Manhattan Streets
Drive an eight-minute taxi shift through Manhattan, in one offline HTML file: passengers hail at the kerb, a GPS routes you over the real street grid, the meter pays by distance, arriving on time and without bumps earns the tip, and a summary card lists every fare. Real street names and landmarks, 200 cars and 340 people in traffic, four times of day, rendered with WebGPU.
Bring the city, the cab and the fare rules, and optionally the look, HUD and shift length. The Skill returns a downloadable Skill (the driving-game project, a customize guide, headless verification) and one playable example shift.
How it works
Your coding agent copies the working game and follows the customize guide: fare rules and destinations, time of day and colour grade, beacons and HUD, the cab, signs and brands.
Routes are planned with A* over the real road network, so every hail and destination is reachable by car.
The shift is played headlessly under the sandbox policy: start, drive, pick up, a bump, the map, drop-off, cockpit camera, night and the end-of-shift card.
Frequently asked questions
Does it call a model or need a key?
No. The Skill gives your agent a working game, a customize guide and verification; the game itself makes no network calls and calls no model.
Why is the demo sandboxed?
The frame has an opaque origin and no network access, so the scene can’t read anything from this page or reach the internet.
Which browsers can run it?
It needs WebGPU: a recent Chrome or Edge on desktop. Where WebGPU is missing the page says so instead of loading the game.
Where does it come from?
It is built on Slop New York by Mr.doob (MIT). We added the taxi shift, our own cab, HUD and colour grade, and replaced real brands. The city, rendering, traffic, people and sound are the original’s work. Map data © OpenStreetMap contributors (ODbL), NYC Open Data and MTA.
Is the game in English only?
Yes. The page is in two languages; the game’s own text is in English.
Does it work on mobile?
Touch controls exist and the layout fits a phone screen in emulation, but it needs WebGPU and a strong GPU; it has not been tested on real phones.
Why is there no online generation?
Building a new scene needs a model key inside the sandbox, which the backend doesn’t support yet. Run the Skill with your own agent for now.